融合GraphRAG与作物模型的极端气候下小麦生产决策智能体

Integrating GraphRAG and Crop Models to Develop Decision-making Agent for Wheat Production under Extreme Climates

摘要:

极端气候事件对冬小麦生产决策时效性和可靠性提出双重挑战,决策过程既需农业领域知识支撑,又需对生育进程等连续过程做出定量判断。通用大语言模型在辅助决策中表现出一定潜力,但仍存在领域知识覆盖不全、数值推理稳定性有限及复杂决策过程可追溯性不足等问题。为此,本文将图检索增强生成（GraphRAG）与作物生长模型CropGrow结合,构建了冬小麦极端气候辅助决策智能体。以农业知识图谱为证据来源,通过模型上下文协议（MCP）将CropGrow封装为数值计算服务,并由协调模块统一完成问题解析、证据检索、作物模型调度和结果校验,使知识证据与机理计算在同一决策流程中联动。在450组极端气候问答和111组田间生育期观测数据上的结果表明,与同基座Qwen3-Max相比,专家核心实体加权召回率由30.69%提升至37.15%,风险相关无支撑表达密度由8.63个/千字降至5.49个/千字;调度CropGrow后,开花期和成熟期预测均方根误差由20.88 d降至5.79 d,决定系数由0.027升至0.896。整体上,智能体的协同调度提高了冬小麦极端灾害情景下知识覆盖和生育期预测能力,为快速形成灾害应对方案提供了可实现的技术路径。

Abstract:

Extreme climate events pose dual challenges to the timeliness and reliability of decision-making in winter wheat production. Effective decision-making requires not only support from domain-specific agricultural knowledge but also quantitative assessment of continuous processes such as crop phenological development. Although general-purpose large language models (LLMs) have demonstrated considerable potential for decision support, their application remains constrained by insufficient coverage of domain knowledge, limited robustness in numerical reasoning, and inadequate traceability in complex decision-making processes. To address these limitations, integrateing graph retrieval-augmented generation (GraphRAG) with the crop model (CropGrow) to develop an intelligent decision-support agent for winter wheat production under extreme climate events. The agent employed an agricultural knowledge graph as its evidence base and encapsulates CropGrow as a numerical computation service through the model context protocol (MCP). A coordination module orchestrated query interpretation, evidence retrieval, crop model invocation, and result validation, thereby enabling knowledge-based evidence and process-based mechanistic computation to operate synergistically within a unified decision-making workflow. Evaluation using 450 extreme-climate question-answer pairs and 111 sets of field-observed phenological data showed that, compared with the same backbone model, Qwen3-Max, the expert-weighted recall of core entities was increased from 30.69% to 37.15%, while the density of unsupported risk-related statements was decreased from 8.63 to 5.49 per 1 000 words. After integrating CropGrow, the root mean square error (RMSE) for flowering and maturity date predictions was decreased from 20.88 days to 5.79 days, while the coefficient of determination (R2) was increased from 0.027 to 0.896. Overall, the coordinated orchestration of the proposed agent improved both domain knowledge coverage and phenological prediction performance under extreme climate disaster scenarios, providing a feasible technical framework for the rapid formulation of disaster-response strategies in winter wheat production.